mistral-large-2512

Overview. Mistral Large 3 is Mistral AI's open-weight flagship: a 675B-parameter Granular sparse Mixture-of-Experts model with 41B active parameters and a native 2.5B-parameter vision encoder, released under Apache-2.0. It ships with a 262,144-token context window and is served via vLLM with the mistral-common tokenizer. Developers who need open, self-hostable weights alongside multimodal (text and image) input and native function calling would reach for this model.

Key strengths. The Granular MoE design keeps active compute at 41B parameters while retaining a 675B-parameter total capacity, which is the main efficiency lever of this release. It accepts text and image input natively through its built-in vision encoder, and Mistral documents native function calling, JSON outputting, and system-prompt adherence as core capabilities. The 262,144-token context window supports long-document and extended multi-turn workflows without external chunking.

When to use it. Use this model for long-document analysis, multimodal tasks that combine text with images, and general instruct workloads where medium task complexity and open weights matter more than top-tier reasoning or coding performance. It's a reasonable default when a team wants an Apache-2.0 licensed, self-hostable model with native tool calling and JSON output for integration into existing pipelines. It is not the model to reach for on heavy agentic or coding-intensive tasks, where independent capability signals in our research place it below top-tier frontier models.

Limitations. Coding and agentic capability signals gathered in research sit below what we'd expect from a top-tier frontier model, so this isn't the right choice for demanding multi-step coding or agent workflows. Mistral has not published a training data cutoff date, and details such as the number of experts per layer and how many are routed per token remain undisclosed. There's also no documented extended-thinking or reasoning-trace mode in this instruct release.

Sovereignty note. Mistral AI is based in France, giving this model no CLOUD Act exposure. Combined with Apache-2.0 open weights, it can be self-hosted for deployments that need direct control over data residency and Sovereign inference.